> The image boorus are longstanding web databases which host large numbers of images which can be ‘tagged’ or labeled with an arbitrary number of textual descriptions; they were developed for and are most popular among fans of anime, who provide detailed annotations.
i.e. it's pulling from a public website which crowdsources the tagging
At present, the most advanced tagger for Danbooru, DeepDanbooru https://www.reddit.com/r/MachineLearning/comments/akbc11/p_t... , still isn't good enough to do annotation by itself but I think that's mostly because no one has really tried.